An AI audit gathers evidence about how a system produced a result. Reviewers may inspect data preparation, model versions, access controls, decision records, infrastructure and the reproducibility of earlier outputs.
High-stakes audits need durable evidence because a decision may be challenged years later. The process should preserve provenance, state and control history while protecting sensitive information and stating the limits of what the audit can establish.
ELI5
An audit is a careful, evidence-based review of how an AI system was built, controlled, and used. It checks the system against stated rules, risks, or requirements rather than relying on promises about how it works.
For example, an audit of an automated lending tool might inspect its data sources, model version, access controls, decision records, and testing results. The reviewer also explains what the available evidence cannot prove.
What evidence does an AI audit use?
It can use data lineage, model and software versions, execution records, access controls, validation results and reproduced outcomes.
Why must AI results be reproducible?
Reproducibility lets investigators verify the conditions and evidence that produced an earlier result rather than trusting a current explanation.
